ε 1 and ε 2 can be either positive or negative, since they are presumably random and symmetrically distributed around zero. When V 12 is large,δ 1 is small (cf. Eqn. 1 ), and as a result the error term ε 1 becomes large relative to δ 1. If δ 1 is small enough, δ 1 + ε 1 will be negative for some measurements (i.e., ε 1 will be negative and larger in magnitude than δ 1). Moreover, the proportion of such measurements will increase asδ 1 decreases, and hence as sap velocity increases. Any algorithm to apply Eqn. 4 must disregard such measurements, because they make the operand of the logarithm negative and thus undefined. But because ε 1 is negative for the discarded points, the remaining (undiscarded) estimates ofδ 1 are positively biased, and hence the resulting estimates of V 12 are negatively biased. This bias increases as the true sap velocity increases, which manifests experimentally as a ”plateau” or ”ceiling” in inferred sap velocity (e.g. Fig. 1), often in the vicinity of 30-50 cm h-1(Flo et al., 2019; Pearsall et al., 2014; Pfautsch et al., 2011). Even if δ 1 is non-negative, noise inδ 1 at high velocities leads to the denominator of the heat ratio approaching zero, with the consequence that resulting estimates of V 12 can fluctuate by orders of magnitude.